A method and system for identifying a time sequence of a crop and a state of a farmland

Abstract

The present application relates to the technical field of cultivated land identification control, and particularly relates to a cultivated land abandoned state time sequence phenology breakpoint identification method and system; the method acquires historical multi-period images, current multi-period images and cultivated land supervision basic data, constructs a map spot level phenology benchmark fingerprint and generates a time sequence image unit; spatial partition identification and cross-view phenology state modeling are performed on the time sequence image unit, a core area, a boundary area, an adjacent ground object interference area and a joint sensitive area are divided, and a map spot level time sequence phenology state sequence is generated; candidate phenology breakpoints are extracted based on the sequence, breakpoint credible potential is calculated, suspected abandoned phenology breakpoints are screened and breakpoint reversibility is judged; then, image slice joints, boundary drift, adjacent ground object interference and contour prediction are combined for reverse rechecking, and finally, an abandoned state identification result and a risk early warning list are output; the scheme can reduce misjudgment caused by image registration error, slice joints and boundary interference, and improve cultivated land abandoned identification accuracy.

Description

Technical Field

[0001] This invention relates to the field of farmland identification and control technology, and in particular to a method and system for identifying time-series phenological breakpoints in farmland abandonment. Background Technology

[0002] Monitoring the status of abandoned farmland is a crucial technical step in farmland protection and farmland outflow analysis. It directly impacts farmland use identification, verification of suspected abandoned farmland plots, supervision of farmland use changes, and the formulation of subsequent protection measures.

[0003] In existing methods for identifying abandoned farmland, platforms typically first identify land use types in single or multiple periods of remote sensing imagery to determine the extent of farmland parcels. Then, they assess whether these parcels exhibit signs of abandonment based on changes in image color, texture, vegetation cover, bare soil exposure, plot boundary changes, or classification results across different periods. This approach can meet routine monitoring needs in scenarios with relatively stable image quality, clear plot boundaries, sufficient sample annotation, and significant changes in farmland use status.

[0004] However, in the actual dynamic monitoring of farmland abandonment, the abandonment state is not a static land type in a single image, but rather a temporal event in which the same farmland patch transitions from a normal phenological process to a non-cultivated or weakly cultivated state across multiple image periods. If we still follow the common land type identification approach, using single-period image slices, pixel classification results, or patch boundary changes as the main criteria, then local visual breaks formed during image segmentation can easily be mistaken for state breaks in the actual phenological process of farmland. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for identifying temporal phenological breakpoints in farmland abandonment. This method constructs a patch-level phenological benchmark fingerprint based on historical normal farming periods, aggregates current image slices into temporal image units according to farmland patches, utilizes spatial partitioning confidence maps to constrain cross-view phenological state modeling, and combines breakpoint confidence potential energy, breakpoint reversibility judgment, and reverse verification of pseudo-breakpoint causes. This achieves the technical effect of reducing misjudgments caused by image slice seams, boundary drift, and interference from adjacent objects, thereby improving the accuracy of farmland abandonment identification.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for identifying time-series phenological breakpoints in the state of abandoned farmland, comprising:

[0008] Based on historical normal farming periods, a patch-level phenological benchmark fingerprint is constructed, and then the current image slices are aggregated according to the cultivated patch to form a time-series image unit.

[0009] Spatial partition confidence maps are constructed based on temporal image units, and then the spatial partition confidence maps are used to constrain cross-view phenological state modeling to form a patch-level temporal phenological state sequence.

[0010] Candidate phenological breakpoints are extracted based on the time-series phenological state sequence at the map patch level. Then, the breakpoint reliability potential energy of the candidate phenological breakpoints is calculated. Subsequently, suspected abandoned phenological breakpoints are screened based on the breakpoint reliability potential energy. Finally, the reversibility of the breakpoints is determined to form the breakpoint type.

[0011] Based on the breakpoint type, a reverse verification of the pseudo-breakpoint causes is performed on the suspected abandoned phenological breakpoints. Then, based on the verification results, the final abandoned status identification result is formed, and a risk warning list is generated.

[0012] Furthermore, a patch-level phenological benchmark fingerprint was constructed based on historical normal farming periods, including:

[0013] For the same cultivated land patch, the occurrence time windows of each phenological stage are extracted from the historical normal cultivation period to generate a historical phenological rhythm timeline;

[0014] Based on the distribution of observation dates corresponding to each phenological stage according to the statistics of historical normal farming periods, the probability window of phenological state is determined.

[0015] Based on the historical phenological rhythm timeline and the phenological state probability window, a patch-level phenological benchmark fingerprint is formed.

[0016] Further, the probability window for phenological states is determined, including:

[0017] For any phenological stage, the distribution of observation dates corresponding to the phenological stage in historical normal farming periods was statistically analyzed;

[0018] Determine the date corresponding to the 10th percentile based on the distribution of observed dates;

[0019] Determine the date corresponding to the 90th percentile based on the distribution of the same observation dates;

[0020] The date range between the 10th percentile and the 90th percentile is defined as the phenological probability window.

[0021] Furthermore, current image slices are aggregated according to cultivated land patches to form temporal image units, including:

[0022] Establish semantic anchoring relationships for each current image slice;

[0023] The farmland patch number corresponding to the current image slice is determined based on the semantic anchoring relationship of the land parcel;

[0024] Aggregate current image slices of the same cultivated land patch using the cultivated land patch number as the primary key;

[0025] The current image slices, arranged and aggregated according to the observation period, form a time-series image unit.

[0026] Furthermore, establish semantic anchoring relationships for land parcels, including:

[0027] Spatially overlay the current image slice with the boundary data of cultivated land patches to determine the number of cultivated land patches covered by the current image slice.

[0028] Determine the relative spatial position of the current image slice in the corresponding cultivated land patch based on the spatial overlay results;

[0029] Determine the interference relationship between adjacent objects based on the features surrounding the boundaries of cultivated land patches;

[0030] The cultivated land parcel numbers are associated with the current image slice, and their relative spatial positions are recorded. Then, the interference relationship with adjacent objects is written to form the semantic anchoring relationship of the land parcel.

[0031] Furthermore, a spatial partitioning confidence map is constructed based on temporal image units, including:

[0032] Extract basic image feature sets from temporal image units;

[0033] Generating fine segmentation results within land parcels based on a set of basic image features;

[0034] The spatial region to which each spatial unit belongs is determined based on the detailed subdivision results within the plot.

[0035] The confidence level of spatial partitioning is determined based on the position of the spatial region relative to the boundary of cultivated land patches.

[0036] Spatial partition confidence scores are arranged according to geographical location to generate a spatial partition confidence map.

[0037] Furthermore, cross-view phenological state modeling includes:

[0038] A spatial feature view constrained by a spatial zoning confidence map is constructed based on the results of fine segmentation within the land parcel.

[0039] The core area is determined based on the spatial zoning confidence map, and the core area image sequence corresponding to the core area is extracted from the temporal image unit;

[0040] Based on the phenological benchmark fingerprint at the patch level, the image sequence of the core area is compared to construct a time-series phenological view;

[0041] When the spatial feature view indicates that the core area has a state interruption during the target observation period and the time-series phenological view determines that the state interruption deviates from the patch-level phenological baseline fingerprint, a suspected phenological anomaly marker is generated.

[0042] Suspected phenological anomalies are marked according to the observation period, forming a time-series phenological state sequence at the patch level.

[0043] Furthermore, the calculation of the reliable potential energy at the breakpoint includes:

[0044] The degree of overlap between the candidate phenological breakpoint and the land parcel boundary buffer zone is obtained, and the degree of overlap between the candidate phenological breakpoint and the image slice seam buffer zone is obtained. Based on this, the intensity of the false breakpoint induction factor is determined.

[0045] The support strength of the true breakpoint is determined based on the sustained significance of the candidate phenological breakpoints within the core region, and then the support strength of the true breakpoints is adjusted based on the sustained deviation of the candidate phenological breakpoints from the patch-level phenological benchmark fingerprint.

[0046] The false breakpoint induction strength is used as the potential increase, and the adjusted true breakpoint support strength is used as the potential decrease to generate the breakpoint credible potential energy.

[0047] Candidate phenological breakpoints whose breakpoint credible potential energy is lower than the preset breakpoint potential energy threshold are identified as suspected abandoned phenological breakpoints.

[0048] Furthermore, determining the breakpoint reversibility leads to the breakpoint type, including:

[0049] Extract a set of reversibility features from the observation periods before and after the suspected phenological breakpoint of abandoned land;

[0050] Determine whether the corresponding cultivated land parcel retains its maintenance features based on the set of reversible features;

[0051] When preserving the maintenance characteristics of the land parcel, suspected abandoned phenological breakpoints are identified as reversible breakpoints;

[0052] When the maintenance characteristics of the land parcel are not preserved and irreversible degradation characteristics appear, the suspected abandoned phenological breakpoints will be determined as irreversible breakpoints.

[0053] The breakpoint type is determined by combining the reversible or irreversible breakpoints obtained from the determination with the abnormal duration period;

[0054] Use the breakpoint type as input for reverse verification of pseudo-breakpoint triggers.

[0055] Furthermore, the reverse verification of spurious breakpoint causes includes:

[0056] First, perform image slice seam verification on suspected abandoned phenological breakpoints;

[0057] Boundary drift verification was performed on the remaining suspected abandoned phenological breakpoints;

[0058] For suspected abandoned phenological breakpoints that are retained again, conduct a review of interference from adjacent objects;

[0059] For the suspected abandoned phenological breakpoints that were retained again, a reverse verification of the profile prediction was performed to obtain the verification results;

[0060] The final identification result of the abandoned land status is formed based on the review results;

[0061] A spatial neighborhood relationship map of cultivated land patches is constructed using cultivated land patches as nodes. Based on the final abandonment status identification results, the abandonment diffusion frontier is identified in the spatial neighborhood relationship map of cultivated land patches, and a risk warning list is generated based on the abandonment diffusion frontier.

[0062] Secondly, the present invention provides a system for identifying time-series phenological breakpoints in the state of abandoned farmland, comprising:

[0063] The image patch benchmark and time series unit generation module is used to construct image patch-level phenological benchmark fingerprints based on historical normal farming periods, and then aggregate the current image slices according to the cultivated map patches to form time series image units.

[0064] The spatial zoning and phenological state modeling module is used to construct a spatial zoning confidence map based on time-series image units, and then use the spatial zoning confidence map to constrain cross-view phenological state modeling to form a patch-level time-series phenological state sequence.

[0065] The module for determining the reliability potential energy and reversibility of breakpoints is used to extract candidate phenological breakpoints based on the time-series phenological state sequence at the map patch level, then calculate the reliability potential energy of the candidate phenological breakpoints, and then screen suspected abandoned phenological breakpoints based on the reliability potential energy of the breakpoints, and finally determine the reversibility of the breakpoints to form the breakpoint type.

[0066] The review output and risk warning module is used to perform reverse review of suspected abandoned phenological breakpoints based on breakpoint type, and then form the final abandoned status identification result based on the review results, and subsequently generate a risk warning list. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0068] Figure 1 This is a flowchart illustrating the overall process of a method for identifying time-series phenological breakpoints in the state of abandoned farmland according to the present invention.

[0069] Figure 2 This is a schematic diagram of the intertemporal boundary drift detection of cultivated land patches in an embodiment of the present invention;

[0070] Figure 3 This is a schematic diagram of individual identification and contour extraction of cultivated land patches in an embodiment of the present invention;

[0071] Figure 4This is a functional module diagram of a time-series phenological breakpoint identification system for abandoned farmland in this invention. Detailed Implementation

[0072] Example 1:

[0073] S1: Construct a phenological baseline fingerprint at the patch level based on historical normal farming periods, and then aggregate the current image slices according to the cultivated patch to form a time-series image unit.

[0074] Farmland abandonment is not a static land type in a single image, but rather a temporal event in which the same farmland patch transitions from a normal phenological process to a weak cultivation state, an uncultivated state, or a continuous abandonment state in consecutive observation periods. If image slices or pixel classification results are directly used as the object of abandonment judgment, image segmentation seams, plot boundary drift, image registration errors, and interference from adjacent objects will be misjudged as actual phenological interruptions. Therefore, this step uses farmland patches as the basic identification object, establishes a phenological reference baseline under historical normal cultivation conditions for each farmland patch, and organizes the current multi-period images into temporal image units that can be continuously processed in subsequent steps.

[0075] Further, step S1 includes:

[0076] S11: Obtain historical multi-period images, current multi-period images, and basic data for farmland monitoring.

[0077] Identifying abandoned land in the monitored area requires simultaneous access to historical normal cultivation reference information, current image information for the period to be identified, and spatial boundary information of cultivated land parcels. If historical normal cultivation reference information is lacking, subsequent steps can only judge anomalies based on a uniform threshold, failing to distinguish between regional differences and the true state of abandoned land. This uniform threshold may include a core area vegetation cover ratio threshold, a bare soil exposure duration threshold, a tillage disturbance frequency threshold, a phenological baseline deviation threshold, and a core area abnormality ratio threshold. For example, the core area vegetation cover ratio threshold is set to 0.30, the bare soil exposure duration threshold to 3 periods, the tillage disturbance frequency threshold to 1 instance, the phenological baseline deviation threshold to 0.40, and the core area abnormality ratio threshold to 0.50. Without cultivated land parcel spatial boundary information, subsequent steps cannot aggregate pixel-level changes into the cultivated land parcel objects required for monitoring.

[0078] Specifically, orthophotos, aerial images, and satellite remote sensing images of the area to be monitored during historical observation periods are obtained from the natural resources image management platform to generate historical multi-period images.

[0079] Specifically, orthophotos, aerial images, and satellite remote sensing images collected according to the same observation calendar within the current identification period are obtained from the natural resources image management platform to generate current multi-period images.

[0080] Specifically, the basic data for farmland supervision is generated by obtaining farmland plot boundary data, permanent basic farmland boundary data, administrative division data, land type sample database data, historical land type identification results, historical farmland use status, historical verified abandoned plot records, and climate anomaly records from the farmland supervision database.

[0081] Specifically, both historical and current multi-period images include image acquisition time, image spatial resolution, image coordinate system, image source type, and image coverage area. The basic data for farmland supervision is stored with the farmland patch number as the primary key.

[0082] S12: Perform geohoming processing on historical and current multi-period images to generate geohomed images.

[0083] Historical and current multi-period images typically originate from different sensors and acquisition batches, resulting in differences in image coordinate systems, spatial resolution, raster grid starting points, and image coverage. Without geohoming processing, the pixel positions of the same cultivated land patch in different observation periods cannot be directly compared, and subsequent changes in phenological conditions will introduce spatial misalignment errors.

[0084] Specifically, using the administrative division boundaries and cultivated land patch boundaries in the basic data for cultivated land supervision as a unified spatial benchmark, coordinate system unified processing is performed on historical and current multi-period images to output images with consistent coordinate systems.

[0085] Specifically, spatial resolution unification processing is performed on coordinate system consistent images. The target spatial resolution is determined based on monitoring accuracy requirements, the effective spatial resolution of images from different periods, and subsequent patch-level identification needs. Images from different sources are then resampled to a unified spatial grid. The target spatial resolution is expressed as the ground sampling distance; for example, the target spatial resolution ranges from 0.5 meters to 2 meters. For images with an original effective spatial resolution lower than the target spatial resolution, the resampling process is only used for spatial grid unification and is not used as a basis for improving the true spatial details of the original images.

[0086] Specifically, image range unification, raster grid alignment, orthorectification, and georegistration are performed on the images after spatial resolution unification to generate multi-period geoconsistent images.

[0087] Specifically, in multi-period geoconsistent imagery, the original geographic coordinates, observation period, image source, cultivated land patch number, and relative spatial position within the cultivated land patch are recorded for each image pixel.

[0088] S13: Records cross-period boundary drift areas based on multi-period geoconsistent imagery and cultivated land patch boundary data.

[0089] Slight spatial misalignments between multiple image periods are typically concentrated near plot edges, patch boundaries, and classification outlines. If edge misalignment is directly used as land cover change input for subsequent steps, these steps will interpret patch boundary drift as a disruption in cultivated conditions, causing suspected fallow phenological breakpoints to cluster towards plot boundaries.

[0090] Specifically, the boundary positions of the same cultivated land patch in different observation periods are overlaid and compared to calculate the spatial offset distance between the boundaries of cultivated land patches in different observation periods.

[0091] Specifically, when the spatial offset distance is greater than the minimum identifiable offset distance but less than or equal to the boundary drift distance threshold, the offset range is recorded as a cross-period boundary drift region. The boundary drift distance threshold is determined based on the ground sampling distance corresponding to the target spatial resolution, specifically twice that ground sampling distance. For example, when the target spatial resolution is 1 meter, the boundary drift distance threshold is 2 meters. The minimum identifiable offset distance is determined based on the ground sampling distance corresponding to the target spatial resolution, specifically taking one ground sampling distance.

[0092] When the spatial offset distance exceeds the boundary drift distance threshold, the corresponding cultivated land patch is marked as a major boundary anomaly patch and recorded in the patch boundary quality anomaly list. Such patches are treated as boundary data quality deficiencies in subsequent breakpoint identification, with the spatial partition confidence of their candidate phenological breakpoints uniformly reduced, and a boundary data quality warning mark is added to the final fallow status identification result. Specifically, the inter-period boundary drift area is written into the inter-period boundary drift record table using cultivated land patch number, observation period combination, boundary drift range, and spatial offset distance as fields.

[0093] like Figure 2 The image shown is a schematic diagram of cross-period boundary drift detection for cultivated land patches. Figure 2 The top left image shows the farmland identification results in March 2025. Figure 2 The bottom left image shows the farmland identification results in April 2025. Figure 2 The right figure shows the difference patches obtained after spatial overlay and boundary comparison of the farmland identification results from the two observation periods mentioned above. Farmland identification results can be generated for the same area under different observation periods, and difference patches can be obtained by comparing results from multiple periods. For the monitoring of farmland abandonment status, such differences may originate from changes in the actual cultivation status, or from image registration errors, patch boundary drift, image resolution differences, or local identification fluctuations.

[0094] Therefore, in step S13, the boundary positions of the same cultivated land patch in different observation periods are overlaid and compared to record the cross-period boundary drift area; and in the subsequent step S18, the cross-period boundary drift marker is written into the time-series image unit. Through this processing, the subsequent steps can distinguish between the actual phenological state transition and the pseudo-change caused by the boundary shift when extracting candidate phenological breakpoints, thereby reducing the risk of misjudgment.

[0095] S14: Construct a patch-level phenological benchmark fingerprint based on historical multi-period images and basic data of cultivated land supervision.

[0096] Different cultivated land patches are affected by topographical conditions, soil conditions, irrigation conditions, and historical farming habits, resulting in inconsistent occurrence times and durations of the greening-up period, vigorous growth period, maturity period, harvest period, plowing period, bare soil period, and revegetation period. If a uniform phenological threshold is used to judge the current anomaly, late-maturing crops, seasonal fallow, and regional differences are easily misjudged as abandoned land.

[0097] Specifically, when selecting historical normal cultivation periods from historical multi-period images, the threshold for historical normal cultivation periods is set to 3 periods, the threshold for phenological state probability window is set to 0.60, the benchmark threshold for phenological response amplitude is set to 0.35, the benchmark threshold for peak vegetation cover in the core area is set to 0.70, the benchmark threshold for bare soil exposure cycle is set to 2 periods, and the benchmark threshold for tillage disturbance frequency is set to once per growing season.

[0098] It should be noted that the historical normal tillage period threshold is used to determine whether a certain cultivated plot has a sample basis for preliminary statistical analysis of historical tillage status. It is mainly used to remove historical periods with obviously insufficient samples or too many abnormal markers. The minimum number of phenological benchmark periods in the subsequent steps is used to determine whether a complete plot-level phenological benchmark fingerprint can be independently constructed. The two have different uses, and the historical normal tillage period threshold can be less than the minimum number of phenological benchmark periods.

[0099] Specifically, historical normal cultivation periods are selected from multiple historical imagery periods. The selection criteria for historical normal cultivation periods are the historical land type identification results, historical farmland use status, and historical verified abandoned land parcel records in the farmland supervision basic data. Observation periods that are not marked as abandoned, occupied by construction, damaged by disasters, or in a non-farmland state are considered historical normal cultivation periods.

[0100] Specifically, for each cultivated plot, the occurrence time window, duration, variation range, and transition speed of adjacent stages of the greening period, vigorous growth period, maturity period, harvest period, plowing period, bare soil period, and revegetation period are extracted from historical normal cultivation periods to generate a historical phenological rhythm timeline.

[0101] Specifically, for each phenological stage of each cultivated plot, the distribution of observation dates of the corresponding phenological stage in historical normal cultivated years is statistically analyzed, and the probability window of phenological state is determined by the equal frequency quantile method; specifically, the 10th percentile and 90th percentile of the date of occurrence of the corresponding phenological stage are calculated, and the date range between the 10th percentile and the 90th percentile is taken as the probability window of phenological state.

[0102] Specifically, based on the peak vegetation cover, bare soil exposure duration, number of tillage disturbances, and revegetation speed of the core area of ​​the cultivated map patch during historical normal cultivation periods, benchmarks for phenological response amplitude, peak vegetation cover in the core area, bare soil exposure cycle, and tillage disturbance frequency are generated.

[0103] The core area refers to the region within a cultivated plot that is far from the plot boundary, unaffected by adjacent objects, and stably reflects the true phenological state of the cultivated crop. Subsequent steps describing stable regions within plots will uniformly use the core area terminology. The boundary area refers to the region located near the boundary of a cultivated plot, easily affected by boundary registration errors, interference from adjacent objects, or plot edge textures. Subsequent steps describing areas affected by plot edges will uniformly use the boundary area terminology. Specifically, the cultivated plot number, historical phenological rhythm timeline, phenological state probability window, phenological response amplitude benchmark, core area vegetation cover peak benchmark, bare soil exposure cycle benchmark, tillage disturbance frequency benchmark, and the most recent normal cultivation state record are combined to generate a plot-level phenological benchmark fingerprint.

[0104] Specifically, the indicators in the phenological baseline fingerprint at the patch level are stored according to a unified statistical standard. Among them, the time window for the occurrence of phenological stages is represented by date or observation day sequence, the duration and bare soil exposure cycle are represented by number of days or observation period, the peak vegetation cover in the core area is represented by vegetation cover ratio or normalized vegetation state index, the frequency of tillage disturbance is represented by the number of occurrences per unit planting cycle, and the phenological response amplitude and revegetation speed are represented by dimensionless normalized index consistent with the corresponding image characteristics.

[0105] Plot-level phenological baseline fingerprints transform the historical normal tillage rhythm of each cultivated plot into a personalized reference baseline, freeing subsequent steps from relying on a uniform threshold across the entire region. For cultivated plots exhibiting low vegetation cover on the same observation date, if the plot-level phenological baseline fingerprint indicates that the corresponding observation date falls within the historical bare soil period or tillage period, subsequent steps will not directly interpret low vegetation cover as abandonment; if the plot-level phenological baseline fingerprint indicates that the corresponding observation date should be within the peak growing season and the peak vegetation cover in the core area is consistently lower than the historical baseline, subsequent steps can attribute the anomaly to deviations in the tillage phenological process.

[0106] S15: Construct patch-level phenological benchmark fingerprints for cultivated land patches with insufficient historical observation periods.

[0107] Newly added cultivated plots, cultivated plots with adjusted boundaries, or cultivated plots with insufficient historical image quality cannot form a stable phenological reference baseline based solely on their own historical normal farming periods. If these cultivated plots are skipped directly, spatial gaps will appear in the monitoring results; if the average phenological baseline for the entire region is applied directly, local differences in topography, soil, and irrigation will lead to subsequent breakpoint judgments.

[0108] Specifically, when the number of historical normal tillage periods for a certain cultivated plot is less than the minimum number of phenological reference periods, the current cultivated plot is marked as a plot with insufficient historical samples. The minimum number of phenological reference periods is determined based on the number of key phenological stages within a complete planting cycle; for example, the minimum number of phenological reference periods is 6 periods.

[0109] The minimum number of phenological benchmark periods is used to determine whether the cultivated plot has historical observation samples covering key stages such as the greening period, vigorous growth period, maturity period, harvest period, tillage period, bare soil period, or revegetation period. If the number of historical normal cultivation periods is lower than this number, a complete plot-level phenological benchmark fingerprint is not directly constructed using the plot's own historical samples, but is instead generated by combining reference samples.

[0110] Specifically, for map patches with insufficient historical samples, candidate similar map patches are selected from the same administrative region. The candidate similar map patches must have no fewer than the number of historical normal farming periods as the minimum phenological reference period.

[0111] Specifically, the similarity of terrain, soil conditions, irrigation conditions, spatial proximity, and historical land use between historically insufficient sample patches and candidate similar patches is calculated. All similarities are uniformly converted into dimensionless values ​​in the range of 0 to 1, and a comprehensive similarity is generated by weighted summation. The larger the comprehensive similarity, the more similar the candidate similar patch is to the historically insufficient sample patch.

[0112] Specifically, the top three candidate similar patches in the comprehensive similarity ranking are selected as reference samples. The historical phenological rhythm time axis, phenological state probability window, phenological response amplitude benchmark, core area vegetation cover peak benchmark, bare soil exposure cycle benchmark, and tillage disturbance frequency benchmark of the reference samples are weighted and averaged according to the comprehensive similarity normalization weight to generate a patch-level phenological benchmark fingerprint for patches with insufficient historical samples.

[0113] S16: Perform image tiling processing on the current multi-period images based on the semantic anchoring relationship of the land parcels to generate an image tile set.

[0114] Subsequent spatial zoning identification and cross-view phenological state modeling require image slices as basic computational input. However, ordinary image slices only contain local image content and do not include the corresponding cultivated land parcel, slice location, or interference areas from adjacent objects. If a semantic anchoring relationship for land parcels is not established at the slicing stage, subsequent steps cannot determine whether anomalies occur in the core area, boundary area, mixed slice coverage area, or image slice coverage area. The semantic anchoring relationship for land parcels refers to the mapping relationship established for each image slice to associate the slice with its corresponding cultivated land parcel, including the cultivated land parcel number covered by the slice, the relative spatial location of the slice within the cultivated land parcel, and the interference relationship between the slice and adjacent objects such as roads, water systems, woodlands, buildings, and ditches.

[0115] Specifically, a sliding window tiling process is performed on the geoconsistent images corresponding to the current multi-period images. The tiling size is determined based on the target spatial resolution, the minimum monitored area of ​​cultivated land parcels, and the requirements for parcel context recognition, so that the ground coverage area corresponding to the image tiling can cover the minimum monitored area and retain parcel boundaries, adjacent features, and core area information. For example, the tiling size is 256 pixels by 256 pixels, and the sliding step size is 128 pixels.

[0116] Specifically, each image slice is spatially overlaid with the boundary data of cultivated land parcels to determine the cultivated land parcel number, cultivated land coverage ratio, and non-cultivated land coverage ratio covered by the image slice.

[0117] Specifically, a slice semantic attribute is established for each image slice. The slice semantic attribute includes slice number, observation period, number of cultivated land patch covered, whether it is completely located inside a single cultivated land patch, whether it crosses the boundary of cultivated land patch, whether it is close to roads, water systems, woodlands, buildings or ditches, whether it is located in the sensitive area of ​​image slice seam, and whether it falls into the cross-period boundary drift area.

[0118] like Figure 3 As shown, this is a schematic diagram of individual identification and contour extraction of cultivated land patches. Based on the artificial intelligence land cover recognition model, cultivated land patches are identified in orthophotos to form plot-level individual segmentation results, and the contours of cultivated land patches are extracted based on the segmentation results.

[0119] The outline of the cultivated land parcels serves as the spatial basis for establishing semantic anchoring relationships between image slices and cultivated land parcel numbers. By spatially overlaying the image slices with the cultivated land parcel outlines, the cultivated land parcel numbers covered by the image slices, the relative spatial position of the slices within the cultivated land parcels, and the interference relationships between the slices and adjacent objects can be determined.

[0120] The results of this type of farmland parcel recognition are not only used to determine the boundaries of farmland parcels, but also for subsequent image tiling processing and the establishment of semantic anchoring relationships between land parcels. Specifically, each image tile is associated with the farmland parcel number it covers, the relative spatial position of the tile within the farmland parcel, and the interference relationship with adjacent objects, enabling subsequent steps to determine whether the anomaly occurs in the core area, the boundary area, the mixed tile coverage area, or the image tile coverage area.

[0121] S17: Based on the image slice set, divide the slices into core slices, boundary slices, mixed slices, and seam-sensitive slices.

[0122] The reliability of land abandonment assessments varies depending on the location of the slice within the image slice set. Slices entirely within cultivated patches are more likely to reflect the true cultivation status of the core area, while slices crossing boundaries or near slice seams are more prone to incorporating road, waterway, woodland shadows, building shadows, and image stitching errors. Differentiating slice types provides a foundation for subsequent spatial zoning confidence maps.

[0123] Specifically, image slices that are entirely located within a single cultivated land patch and whose distance from the cultivated land patch boundary is greater than the core area distance threshold are marked as core slices. The core area distance threshold is determined based on the ground sampling distance corresponding to the target spatial resolution and the average boundary registration error of the cultivated land patch. For example, when the target spatial resolution is 1 meter and the average boundary registration error of the cultivated land patch is 1 meter, the core area distance threshold is 3 meters.

[0124] Specifically, image slices that intersect with the boundaries of cultivated land patches are marked as boundary slices.

[0125] Specifically, image slices that simultaneously cover both cultivated and non-cultivated land features, with the non-cultivated land coverage ratio exceeding a non-cultivated land mixing threshold, are marked as mixed slices. The non-cultivated land mixing threshold is determined based on the statistical distribution of mixed slices in the land use sample library; for example, the non-cultivated land mixing threshold is set to 0.2.

[0126] Specifically, image slices located near the edges of image slices, splicing edges, or segmentation seams are marked as seam-sensitive slices.

[0127] S18: Aggregate image slice sets, inter-period boundary drift records, and patch-level phenological benchmark fingerprints using cultivated land patch numbers as the primary key to generate time-series image units.

[0128] The subsequent step S2 requires comparing the changes in phenological states at different observation periods within the same cultivated land patch. Therefore, the input cannot be unordered image slices, but should be a structured object aggregated by cultivated land patch number, arranged by observation period, and carrying both patch-level phenological baseline fingerprints and spatial interference markers.

[0129] Specifically, using the cultivated land patch number as the primary key, the core slice, boundary slice, mixed slice, and seam sensitive slice of the same cultivated land patch under different observation periods are retrieved from the image slice set.

[0130] Specifically, the cross-period boundary drift record table corresponding to the same cultivated land patch, the phenological benchmark fingerprint of the patch, the historical identification results of the patch, and the set of multiple period image fragments are written into the same data object.

[0131] Specifically, the core area image sequence, boundary area image sequence, adjacent object interference area image sequence, and seam sensitive area image sequence of the same cultivated land parcel are arranged in chronological order according to the observation period to generate a time-series image unit.

[0132] Specifically, the temporal image unit includes cultivated land patch number, multi-period image fragment set, core area image sequence, boundary area image sequence, adjacent ground object interference area image sequence, joint sensitive area image sequence, cross-period boundary drift marker, patch history identification results, and patch-level phenological benchmark fingerprint of the corresponding patch.

[0133] S2: Construct a spatial partition confidence map based on time-series image units, and then use the spatial partition confidence map to constrain the cross-view phenological state modeling to form a patch-level time-series phenological state sequence.

[0134] While the temporal image units have established data organization relationships at the cultivated land parcel dimension, the land cover status, spatial location confidence, and deviation from the parcel-level phenological baseline fingerprint for each observation period have not yet been transformed into temporal state nodes that can be used for breakpoint identification. If breakpoint identification is directly performed on the temporal image units, boundary anomalies, seam anomalies, and true anomalies in the core area cannot be distinguished. Therefore, this step transforms image information into parcel-level temporal phenological state sequences through spatial partitioning identification and cross-view phenological state modeling.

[0135] Further, step S2 includes:

[0136] S21: Extract the edge, texture and local disturbance features of the land parcels from the time-series image units to generate a basic image feature set.

[0137] Normal tillage conditions in cultivated land parcels typically manifest as phased changes in vegetation cover, tillage texture, and bare soil exposure. However, road edges, ditches, waterways, woodland shadows, building shadows, and irrigated wet areas exhibit localized appearance variations unrelated to tillage conditions. The base image feature set provides computational input for subsequent coarse segmentation, fine segmentation, and spatial partitioning confidence modeling.

[0138] Specifically, color space transformation is performed on each observation period image in the time-series image unit to generate a color feature map.

[0139] Specifically, edge enhancement, texture gradient extraction, shadow region identification, local contrast analysis, bare soil exposure detection, and morphological processing are performed on the color feature map to generate image edge features, texture continuity features, shadow perturbation features, bare soil exposure features, and vegetation cover features.

[0140] Specifically, feature statistics are performed on the interior and periphery of cultivated land patches, and field ridges, ditches, water system edges, woodland shadows, road edges, building shadows, irrigation wet areas, and image slice seam areas are marked as boundary sensitive areas or adjacent object interference areas.

[0141] The term "adjacent object interference zone" refers to the interference area formed when cultivated land patches are adjacent to roads, water systems, woodlands, buildings, ditches, or other non-cultivated land features, due to the influence of shadows, moisture, texture extension, vegetation expansion, or edge pixel blending. The term "adjacent object interference zone" will be used consistently in subsequent steps.

[0142] Specifically, image edge features, texture continuity features, shadow perturbation features, bare soil exposure features, and vegetation cover features are combined to generate a basic image feature set; and boundary sensitive areas and adjacent object interference areas are used as region markers and associated with the basic image feature set for storage.

[0143] S22: Perform coarse and fine segmentation based on the basic image feature set to generate fine segmentation results within the plot.

[0144] Subsequent spatial zoning confidence maps need to identify which areas within cultivated land parcels belong to the core area, boundary area, field ridge area, ditch area, bare soil area, vegetation-covered area, waterlogged area, weed-covered area, and area disturbed by adjacent objects. Basic image feature sets alone cannot directly support regional confidence assignment; further land use segmentation and refinement of the internal structure of each parcel are required.

[0145] Specifically, multi-period image slices from a time-series image unit are input into a coarse segmentation model to identify cultivated land, forest land, water systems, roads, buildings, bare land, and other land types, generating a coarsely segmented land type map. The coarse segmentation model employs a semantic segmentation network based on a U-Net structure. The U-Net structure is an encoder-decoder semantic segmentation network structure used to extract multi-scale image features and recover pixel-level land type segmentation results. The training samples for the semantic segmentation network come from a land type sample library in the cultivated land monitoring basic data, and the training labels include cultivated land, forest land, water systems, roads, buildings, bare land, and other land types.

[0146] Specifically, the coarse-divided land use map is used to determine the candidate range of cultivated land, the interference zone of adjacent features, and the spatial adjacency relationship between cultivated land patches and surrounding features.

[0147] Specifically, the candidate cultivated land area from the coarse-segmented land use map is input into the fine-segmentation model to generate fine-grained segmentation results within each plot. The fine-segmentation model employs the DeepLabv3+ semantic segmentation network. This network is used to further extract internal structural features of plots within the candidate cultivated land area and generate fine-grained segmentation results. The DeepLabv3+ semantic segmentation network is trained using samples from the cultivated land sample library that contain annotations of the internal structure of plots.

[0148] S23: Construct a spatial zoning confidence map based on the results of fine segmentation within the plot, inter-period boundary drift areas, and slice type classification.

[0149] The detailed segmentation results within the plots describe the land cover categories, but the reliability of the abandonment judgment varies in different spatial areas. Areas located in the core area and with stable identification over multiple periods are more suitable as the basis for judging the actual phenological interruption; areas located in areas with roads, water systems, woodlands, building shadows, image slice seams, or cross-period boundary drift are more likely to form false breakpoints.

[0150] Specifically, using the cultivated land parcel number as the primary key, the fine segmentation results within the parcel, the cross-period boundary drift area, the core slice, the boundary slice, the mixed slice, and the seam sensitive slice are spatially overlaid to generate a spatial zoning base map.

[0151] Specifically, the region to which each spatial unit belongs is determined based on the spatial zoning baseline map, and the spatial zoning confidence level is determined by combining the spatial relationship between the spatial unit and the boundaries of cultivated land patches, image tile seams, and inter-period boundary drift areas. Spatial units located in the core area and far from the interference areas have higher spatial zoning confidence levels; spatial units located in the boundary area, adjacent object interference area, sensitive area of ​​image tile seam, or inter-period boundary drift area have lower spatial zoning confidence levels. The spatial zoning confidence levels are uniformly mapped to the interval [0,1], with larger values ​​indicating that the corresponding spatial unit is more suitable as a basis for judging the state of abandoned land.

[0152] Specifically, the confidence level of spatial partitioning ranges from [0,1]. The larger the value, the more suitable the corresponding spatial unit is as a basis for abandonment judgment. The smaller the value, the more likely the corresponding spatial unit is to be affected by boundaries, seams, adjacent objects, or registration drift.

[0153] Specifically, the spatial partition confidence scores of all spatial units are arranged according to their geographical location to generate a spatial partition confidence map.

[0154] S24: Construct spatial feature views and temporal phenological views, and perform cross-view... Figure 1 Consistency constraints are applied to generate a time-series phenological state sequence at the patch level.

[0155] Anomalies in the appearance of land features in a single image do not necessarily indicate a state of abandonment, and deviations in phenology across consecutive observation periods also need to be judged in conjunction with the spatial regional reliability. Spatial feature views can identify the structural state of land parcels in each image period, while time-series phenological views can identify continuous deviations of the same cultivated land parcel from the parcel-level phenological baseline fingerprint. The combined constraints of these two types of views can reduce misjudgments from a single view.

[0156] Specifically, the spatial feature view takes the basic image feature set, the fine segmentation results of the plots, and the spatial zoning confidence map as inputs, learns the spatial structure of cultivated land parcels in a single period of imagery, and outputs the feature status representation of the core area, boundary area, and adjacent feature interference area in each observation period.

[0157] Specifically, the time-series phenological view takes the core area image sequence, the patch-level phenological benchmark fingerprint, and the observation period as input. It compares the current observation sequence with the historical phenological rhythm time axis, phenological state probability window, phenological response amplitude benchmark, core area vegetation cover peak benchmark, bare soil exposure cycle benchmark, and tillage disturbance frequency benchmark, and outputs the deviation amplitude of the patch-level phenological benchmark fingerprint.

[0158] Specifically, a first pseudo-label is generated by inputting a weakly enhanced image into a spatial ground view; a second pseudo-label is generated by inputting a weakly enhanced temporal sample into a temporal phenological view. The weakly enhanced image is obtained through brightness perturbation, contrast perturbation, and slight rotation, and the weakly enhanced temporal sample is obtained by maintaining the order of observation periods and adding low-amplitude noise.

[0159] Specifically, the first pseudo-label is used to supervise the output of the temporal phenological view on the strongly enhanced temporal sample, and the second pseudo-label is used to supervise the output of the spatial feature view on the perturbed image sample. The strongly enhanced temporal sample is obtained through simulation of random occlusion, strong brightness perturbation, and temporal local missing measurements, and the perturbed image sample is obtained through spatial cropping, color perturbation, and texture perturbation.

[0160] Specifically, when both the spatial ground feature view and the temporal phenological view indicate a state interruption in the same observation period in the core area, and the state interruption deviates from the patch-level phenological benchmark fingerprint, a suspected phenological anomaly marker is generated.

[0161] Specifically, when anomalies are found only in the boundary area, joint area, or adjacent object interference area in the spatial feature view and the temporal phenology view, but no synchronous anomalies are found in the core area, the anomalies are marked as boundary-induced anomalies or joint-induced anomalies.

[0162] Specifically, the spatial feature view characteristics, temporal phenological view characteristics, spatial zoning confidence map, and deviation magnitude of the phenological baseline fingerprint of each cultivated land patch in each observation period are fused to generate a temporal phenological state sequence at the patch level.

[0163] Specifically, the map-level temporal phenological state sequence includes multiple temporal state nodes. Each temporal state node includes the observation period, phenological state of the core area, phenological state of the boundary area, state of the adjacent ground object interference area, land type identification stability, spatial zoning confidence, and cross-view. Figure 1 Consistency markers, cross-view consistency assessment, deviation of phenological baseline fingerprint at the patch level, boundary-induced anomaly markers, and suspected phenological anomaly markers.

[0164] Among them, the stability of land use identification is determined based on the consistency of land use identification results in adjacent observation periods for the spatial area corresponding to the same cultivated land patch; the larger the value, the more stable the land use identification results across multiple periods; cross-view Figure 1 Consistency markers are used to indicate whether spatial feature views and temporal phenological views provide consistent status interruption judgments for the core area during the same observation period. The consistency degree across views is determined based on the consistency of judgments regarding the phenological state and anomalies of the core area by spatial feature views and temporal phenological views; a larger value indicates greater consistency between the two types of views. The deviation magnitude of the patch-level phenological benchmark fingerprint is determined based on the deviation of the current observation period's phenological state, peak vegetation cover in the core area, bare soil exposure cycle, and tillage disturbance frequency from the corresponding items in the patch-level phenological benchmark fingerprint; a larger value indicates a more significant deviation of the current phenological state from historical normal cultivation conditions. The land cover identification stability, spatial zoning confidence, cross-view judgment consistency degree, and patch-level phenological benchmark fingerprint deviation magnitude are all dimensionless indicators and uniformly mapped to the [0,1] interval.

[0165] Spatial feature views and temporal phenological views form complementary constraints. Spatial feature views can identify single-period spatial structural anomalies using plot shape, field texture, bare soil distribution, and adjacency relationships. However, when spatial feature views lack historical phenological references, they can easily interpret seasonal tillage or short-term bare soil as anomalies. Temporal phenological views can determine whether the current observation sequence deviates from the patch-level phenological baseline fingerprint. However, when temporal phenological views lack spatial location constraints, they can easily include boundary drift and seam disturbances in discontinuity judgments. When both types of views simultaneously provide state discontinuity results in the core area, the anomaly simultaneously satisfies spatial authenticity and temporal persistence. When both types of views only provide anomalies in the boundary area or seam area, the anomaly is restricted to a spurious discontinuity trigger. This cross-view Figure 1 Consistency constraints enable subsequent breakpoint identification to obtain inputs that simultaneously include spatial credibility and phenological continuity.

[0166] To further identify farmland abandonment, spatial feature views and temporal phenological views are introduced. Spatial feature views are used to determine the spatial location and feature state of anomalies within farmland patches, while temporal phenological views are used to determine whether the current observation sequence deviates from the patch-level phenological baseline fingerprint. A suspected phenological anomaly marker is generated only when both types of views indicate a state interruption in the core area for the same observation period, and this state interruption deviates from the patch-level phenological baseline fingerprint. This avoids directly equating general image changes with phenological breakpoints in farmland abandonment.

[0167] S3: Based on the time-series phenological state sequence of the map patch, candidate phenological breakpoints are extracted, and then the breakpoint reliability potential energy of the candidate phenological breakpoints is calculated. Subsequently, suspected abandoned phenological breakpoints are screened based on the breakpoint reliability potential energy, and finally the breakpoint reversibility is determined to form the breakpoint type.

[0168] The plot-level temporal phenological state sequence has described the phenological state of each cultivated plot in different observation periods, but the judgment of fallow land cannot be completed based on a single anomaly node. True fallow land breakpoints are usually manifested as continuous anomalies in the core area, persistent deviations from the plot-level phenological baseline fingerprint, and discrepancies between the spatial feature view and the temporal phenological view. Figure 1 The anomaly is consistent and not caused by boundaries, seams, adjacent objects, or registration drift. This step incorporates true breakpoint support factors and false breakpoint causes into the same judgment framework through breakpoint reliability potential energy, and further determines breakpoint reversibility.

[0169] Further, step S3 includes:

[0170] S31: Traverse the temporal phenological state sequence at the patch level and extract candidate phenological breakpoints.

[0171] If anomalies are detected only in a single observation period and a fallow status is output, short-term cloud cover, irrigation waterlogging, bare soil after harvest, and seasonal fallow may all be misjudged. Candidate phenological breakpoints should reflect state transitions in adjacent or consecutive observation periods, and the state transitions need to deviate from the phenological baseline fingerprint at the patch level.

[0172] Specifically, the phenological state sequence of each cultivated plot is traversed in the order of observation period, and the phenological state of the core area in adjacent phenological state nodes is compared.

[0173] Specifically, when the phenological state of the core area changes from normal cultivation state to weak cultivation state, no cultivation disturbance state, long-term bare soil state, continuous weed cover state, or natural vegetation succession state, candidate state transition records are generated.

[0174] Specifically, the observation date, observation period, and identified current phenological stage corresponding to the candidate state transition record are compared with the phenological state probability window in the patch-level phenological benchmark fingerprint. The phenological state probability window is the statistical time range of the dates that the corresponding phenological stage appears in historical normal farming periods, expressed in date or observation day sequence.

[0175] When the current phenological stage corresponding to the current observation date does not fall within the phenological state probability window of that phenological stage, or when the current observation date falls within the time window of the historical patch-level phenological baseline fingerprint that should be characterized by normal cultivation, peak growth period, or revegetation period, but the current phenological state of the core area is characterized by weak cultivation, no cultivation disturbance, long-term bare soil, continuous weed coverage, or natural vegetation succession, the candidate state transition record will be marked as a candidate phenological breakpoint.

[0176] The observation date and the probability window for phenological states are both represented by date or observation day sequence, and the two have the same dimensions. The current phenological state, as the result of the category judgment, is not directly compared with the date window, but is compared with the expected phenological stage under the observation date.

[0177] Specifically, when the current phenological state does not fall outside the phenological state probability window, but the current index values ​​of the phenological response amplitude, peak vegetation cover in the core area, frequency of tillage disturbance, or bare soil exposure cycle exceed the normal range of the corresponding index in the patch-level phenological benchmark fingerprint, the candidate state transition record is marked as a candidate phenological breakpoint. The normal range is determined based on the historical mean and standard deviation of the corresponding index in historical normal tillage periods, specifically the range formed by adding or subtracting twice the standard deviation from the historical mean; for indices that only have unilateral abnormal meaning, the boundary value in the corresponding direction is used as the judgment basis.

[0178] S32: Candidate anomalies are induced by filtering boundaries based on spatial partition confidence maps and pseudo-breakpoint causes.

[0179] Candidate phenological breakpoints may still originate from boundary areas, image slice seam areas, mixed slice coverage areas, adjacent object interference areas, or cross-period boundary drift areas. If the causes of false breakpoints are not filtered out, a large number of candidate anomalies unrelated to actual abandonment will be mixed into the subsequent calculation of breakpoint reliability potential energy, increasing the burden of manual verification.

[0180] Specifically, the abnormal spatial range corresponding to the candidate phenological breakpoint is overlaid with the spatial partition confidence map, and the area ratio of the abnormal spatial range in the core area, boundary area, image slice seam area, mixed slice coverage area, adjacent object interference area, and inter-period boundary drift area is calculated.

[0181] Specifically, when the abnormal spatial range only occurs in the boundary area, image slice seam area, mixed slice coverage area, adjacent ground object interference area, or cross-period boundary drift area, and there is no synchronous phenological interruption in the core area, no real candidate phenological breakpoint is generated, but a pseudo breakpoint candidate record is generated.

[0182] Specifically, the candidate records for false breakpoints include image slice seam induced anomalies, boundary drift induced anomalies, adjacent object induced anomalies, shadow induced anomalies, land cover segmentation instability anomalies, and contour correction induced anomalies.

[0183] S33: Calculate the breakpoint reliability potential energy of candidate phenological breakpoints and generate a record of breakpoint reliability potential energy.

[0184] The next step requires classifying candidate phenological breakpoints into suspected fallow phenological breakpoints or pseudo-breakpoints with unverified potential energy. Therefore, a dimensionless comprehensive discrimination index is needed that can simultaneously characterize the degree of support for true fallow and the strength of pseudo-breakpoint inducing factors. This paper refers to this index as the breakpoint credibility potential energy. It should be noted that the breakpoint credibility potential energy adopts an inverse credibility design; that is, the lower the breakpoint credibility potential energy, the closer the candidate phenological breakpoint is to true fallow or persistent weak cultivation characteristics; the higher the breakpoint credibility potential energy, the more likely the candidate phenological breakpoint originates from pseudo-breakpoint inducing factors such as boundaries, image slice seams, adjacent objects, registration drift, or model instability.

[0185] Specifically, based on the pseudo-breakpoint candidate records and spatial partitioning results obtained in step S32, within the observation window after the candidate phenological breakpoints, the total number of abnormal evidence, the number of pseudo-breakpoint evidence, and the number of real breakpoint evidence are counted respectively.

[0186] Specifically, the reliable potential energy at the breakpoint is calculated using the following formula:

[0187]

[0188] In the formula, This is the total number of abnormal pixel periods determined by step S31 within 5 valid observation periods after the breakpoint; The number of pixel periods that fall into the pseudo-breakpoint inducement areas such as boundaries, seams, adjacent objects or shadows, and inter-period boundary drift recorded in step S32; The number of pixel periods located in the core area that deviate from the phenological baseline fingerprint for at least two consecutive periods within the five periods is defined as the number of periods. The proportion of false breakpoint evidence in all anomalous evidence represents the intensity of the false breakpoint induction, while the corresponding proportion of true breakpoint evidence represents the strength of true breakpoint support. The boundary buffer radius is the sum of 2s and the average boundary registration error, the seam buffer radius is 2s, the adjacent object buffer radius is 3s, and the boundary drift distance threshold is 2s, where s is the ground sampling distance. Clouds, shadows, and invalid values ​​are first masked, and linear interpolation is performed when there are no more than two consecutive missing periods; if there are more than two missing periods, it is marked as insufficient data.

[0189] When it increases, An increase indicates a higher probability of a false breakpoint; When it increases, A decrease indicates an increased likelihood of actual land abandonment. Located in the [-1,1] interval, and without any weights to be fitted, no training data or weight verification is required. The pseudo-breakpoints referred to in this paper are candidate breakpoints caused by factors such as boundaries, seams, adjacent objects, or intertemporal drift.

[0190] Specifically, the number of cultivated land patches, the occurrence period of candidate phenological breakpoints, and , , The breakpoint credible potential energy is written into the breakpoint credible potential energy record.

[0191] The formula uses the pseudo-breakpoint evidence obtained in the previous step as the increasing potential and the real breakpoint evidence that the core area continuously deviates from the phenological benchmark fingerprint of the patch level as the decreasing potential, thereby converting the pseudo-breakpoint filtering result of step S32 into the breakpoint credible potential energy that can be directly compared in step S34.

[0192] S34: Filter suspected abandoned phenological breakpoints based on the breakpoint reliable potential energy record, and generate breakpoint interpretation chains.

[0193] The breakpoint reliability energy record provides a quantitative ranking of candidate phenological breakpoints, but regulatory applications require interpretable breakpoint results, not just numerical outputs. The breakpoint interpretation chain can explain the occurrence period, spatial location, duration, abnormal area ratio of the core area, and the results of excluding false breakpoint causes of suspected abandoned phenological breakpoints, providing a basis for manual verification and administrative action.

[0194] Specifically, the preset breakpoint potential energy threshold is set to 0; When less than 0, Greater than It was identified as a suspected breakpoint in the phenological record of abandoned land. When the potential energy is greater than or equal to 0, it is identified as a pseudo-breakpoint with potential energy to be verified. Therefore, the lower the potential energy of the breakpoint, the higher the confidence level of the candidate breakpoint.

[0195] Specifically, candidate phenological breakpoints whose reliable breakpoint potential energy is higher than or equal to a preset breakpoint potential energy threshold but still exhibits abnormal changes are identified as pseudo-breakpoints requiring potential energy verification. These pseudo-breakpoints refer to candidate breakpoints that have not yet met the criteria for suspected abandoned phenological breakpoints but still show signs of image or phenological anomalies, requiring further confirmation by the image processing module, the patch boundary maintenance module, or a manual verification process. Pseudo-breakpoints requiring potential energy verification are not directly output as abandoned results and do not proceed to the breakpoint reversibility judgment process in step S35.

[0196] Therefore, candidate phenological breakpoints are first divided into suspected fallow phenological breakpoints and pseudo-breakpoints with potential energy to be verified based on the reliable potential energy of the breakpoints. Among them, suspected fallow phenological breakpoints enter the subsequent breakpoint reversibility judgment process, while pseudo-breakpoints with potential energy to be verified enter the pseudo-breakpoint verification output process and do not enter the breakpoint reversibility judgment process.

[0197] Specifically, for each suspected phenological breakpoint, a breakpoint interpretation chain is generated. The breakpoint interpretation chain includes the breakpoint occurrence period, whether the breakpoint is located in the core area, the duration of the breakpoint, the proportion of abnormal area in the core area, the deviation of the phenological benchmark fingerprint at the patch level, and the cross-viewing distance. Figure 1 Consistency results, whether image slice seams were avoided, whether boundary drift areas were avoided, whether interference areas from adjacent objects were avoided, and the results of eliminating false breakpoint causes.

[0198] S35: Perform a breakpoint reversibility judgment on suspected abandoned phenological breakpoints to form a breakpoint type.

[0199] A suspected phenological breakpoint in land abandonment is not entirely equivalent to continuous abandonment. Seasonal fallow, short-term weak cultivation, and recoverable weak cultivation can also manifest as phenological breaks. If the reversibility of the breakpoint is not determined, regulatory results will confuse short-term recoverable states with continuous abandonment states, affecting the priority of subsequent manual verification and administrative handling strategies.

[0200] Specifically, the residual traces of cultivation, soil structure integrity, degree of maintenance of field ridge texture, integrity of irrigation facilities, regularity of plot boundaries, and duration of short-term bare soil exposure are extracted from the observation periods before and after the suspected fallow phenological breakpoint to generate a set of reversible features.

[0201] Specifically, when the set of reversibility features shows that cultivated plots still retain soil and plot maintenance features from before normal cultivation, including residual cultivation traces, intact soil structure, maintained field ridge texture, undamaged irrigation facilities, regularity of plot boundaries, and short-term exposure of bare soil, suspected fallow phenological breakpoints are identified as reversible breakpoints.

[0202] Specifically, when perennial weeds invade, shrubs sprout, field ridges break, ditches become silted up, surface micro-topography degrades, cultivation textures disappear, and vegetation cover changes from crop type to natural weed type in the cultivated plots corresponding to suspected fallow phenological breakpoints, the suspected fallow phenological breakpoints are identified as irreversible breakpoints.

[0203] Specifically, based on the credible potential energy of the breakpoint, the result of the breakpoint reversibility judgment, the duration of the abnormality, and whether there are signs of subsequent recultivation, suspected fallow phenological breakpoints are divided into suspected continuous fallow breakpoints, suspected seasonal fallow breakpoints, and weakly recoverable tillage breakpoints. Specifically, when a suspected fallow phenological breakpoint is determined to be an irreversible breakpoint, and the core area continuously deviates from the map-level phenological baseline fingerprint, the cultivated texture disappears, or the natural weed vegetation continuously strengthens, it is classified as a suspected continuous fallow breakpoint; when a suspected fallow phenological breakpoint is determined to be a reversible breakpoint, and the occurrence of anomalies matches the local fallow, crop rotation, or agricultural intermission periods, and the field ridge texture, soil structure, and conditions for recultivation are preserved, it is classified as a suspected seasonal fallow breakpoint; when a suspected fallow phenological breakpoint is determined to be a reversible breakpoint, but the duration of the anomaly exceeds the regular fallow period or is characterized by a decrease in the frequency of tillage disturbance and a decrease in the peak vegetation cover in the core area without showing irreversible degradation characteristics, it is classified as a recoverable weak cultivated breakpoint.

[0204] S4: Based on the breakpoint type, perform reverse verification of the pseudo-breakpoint causes for suspected abandoned phenological breakpoints, and then form the final abandoned status identification result based on the verification results, and subsequently generate a risk warning list.

[0205] After step S3, the suspected abandoned phenological breakpoints have obtained the breakpoint credibility potential energy, breakpoint interpretation chain, and reversibility judgment results. However, the final regulatory output still needs to exclude residual pseudo-breakpoints caused by image slice seams, boundary drift, interference from adjacent objects, and contour correction, and transform the abandoned land identification results into an executable priority list for manual verification and a risk warning list. This step correlates the breakpoint identification results with pseudo-breakpoint verification, status output, and spatial diffusion risk analysis to generate the final map-level abandoned land status identification results and risk warning results.

[0206] Further, step S4 includes:

[0207] S41: Perform image slice seam verification, boundary drift verification, adjacent object interference verification, and contour prediction reverse verification on suspected abandoned phenological breakpoints.

[0208] Step S32 has eliminated candidate anomalies that only occur in the boundary area, seam area, or adjacent object interference area based on area ratio. However, there are still candidate breakpoints that cross both the core area and the boundary area and have not been completely eliminated in the spatial zoning confidence map. Step S41 performs a precision fallback for these residual cases to verify whether the suspected abandoned phenological breakpoints depend on the existence of spurious breakpoints. The verification operation can verify whether the suspected abandoned phenological breakpoints depend on the existence of spurious breakpoints from both the spatial distribution pattern and the continuity of the core area.

[0209] Specifically, it is determined whether suspected fallow phenological breakpoints are distributed along the seams of image slices; when the spatial distribution of suspected fallow phenological breakpoints is consistent with the direction of the image slice seams, and the two sides of the breakpoint belong to the core area of ​​the same cultivated land patch, the suspected fallow phenological breakpoints are marked as image slice seam induced breakpoints, and the fallow credibility is reduced; when suspected fallow phenological breakpoints span multiple image slices and exist continuously in the core area, the suspected fallow phenological breakpoints are retained.

[0210] Specifically, the boundary positions of the same cultivated land patch in different observation periods are compared to identify inter-period boundary drift areas. When the suspected fallow phenological breakpoint is mainly located within the inter-period boundary drift area and there are no synchronous phenological anomalies in the core area, the suspected fallow phenological breakpoint is marked as a boundary drift-induced breakpoint. When the suspected fallow phenological breakpoint is far away from the inter-period boundary drift area and continues to appear in the core area, the fallow judgment of the suspected fallow phenological breakpoint is retained.

[0211] Specifically, it is determined whether the suspected phenological breakpoint is located at the boundary between cultivated land and forest land, water system, road, building, ditch or other land features; when the suspected phenological breakpoint highly overlaps with the adjacent land feature interference area, and the direction of image change is consistent with the shadow, moisture, texture, vegetation expansion or road edge characteristics of the adjacent land feature, the suspected phenological breakpoint is marked as an adjacent land feature induced breakpoint; when the suspected phenological breakpoint does not spatially overlap with the adjacent land feature interference area, and the phenological state of the core area is continuously interrupted, the abandonment judgment of the suspected phenological breakpoint is retained.

[0212] Specifically, based on the results of contour rectangularization, boundary prediction, or patch boundary correction, the boundaries of cultivated patches before and after the occurrence of suspected fallow phenological breakpoints are predicted and compared. When the suspected fallow phenological breakpoints are mainly caused by the expansion, contraction, smoothing, or rectangularization correction of cultivated patch ranges before and after boundary prediction, and there is no synchronous phenological breakpoint in the core area, the suspected fallow phenological breakpoints are marked as contour correction-induced breakpoints. When the suspected fallow phenological breakpoints still exist stably in the core area before and after contour prediction processing, it is confirmed that the suspected fallow phenological breakpoints do not depend on the geometric boundary correction results.

[0213] S42: Perform reversible review linkage processing based on the review results and the breakpoint reversibility judgment results.

[0214] The verification results are used to eliminate false breakpoint causes, while the breakpoint reversibility judgment results are used to distinguish between continuous fallow, seasonal fallow, and weakly recoverable tillage. If the two types of results are not linked, the system may output results that are spatially reliable but have inaccurate state types, or output results with reasonable state types but affected by false breakpoint causes.

[0215] Specifically, when a suspected fallow phenological breakpoint is verified through step S41 and determined to be an irreversible breakpoint, the corresponding cultivated land patch is marked as a suspected continuous fallow patch.

[0216] Specifically, when a suspected fallow phenological breakpoint is verified through step S41 and determined to be a reversible breakpoint, the corresponding cultivated land patch is marked as a suspected seasonal fallow patch or a weakly recoverable cultivated land patch, and the corresponding cultivated land patch is written into the next phase of the image automatic tracking queue.

[0217] Specifically, when a cultivated plot in the next image automatic tracking queue shows a recultivation signal in the next observation period, the fallow marker is removed. The recultivation signal includes the probability window of the phenological state corresponding to the core area vegetation cover entering the plot-level phenological benchmark fingerprint, the frequency of tillage disturbance returning to the tillage disturbance frequency benchmark range, the field texture returning to the tillage texture, and the bare soil exposure period falling back to the bare soil exposure period benchmark range.

[0218] Specifically, when the cultivated land patches in the next phase of the automatic image tracking queue continue to be interrupted in the next observation period and the reversible characteristics disappear, the corresponding cultivated land patches will be upgraded to suspected patches of continuous abandonment.

[0219] S43: Construct a spatial neighborhood relationship map of cultivated land patches and generate a list of abandoned land risk warnings.

[0220] Abandonment is often spatially related to the adjacency of plots, the obstruction of roads and waterways, and the connectivity of the terrain. In areas where suspected persistent abandonment has already occurred, the risk of subsequent abandonment is higher if other cultivated plots simultaneously show a decline in phenological activity. A risk warning list can expand supervision from post-event verification to pre-event intervention.

[0221] Specifically, a spatial neighborhood relationship graph of cultivated land parcels is constructed, using cultivated land parcels as nodes and the spatial adjacency relationships or distances less than a preset neighborhood radius between parcels as edges. The preset neighborhood radius is determined based on the average cultivated land parcel area of ​​the area to be monitored. The average cultivated land parcel area is calculated based on the parcel area under a unified projection coordinate system, and specifically, the preset neighborhood radius is taken as three times the equivalent circle radius of the average cultivated land parcel.

[0222] Specifically, the edge weight calculation comprehensively considers the distance between map features, the length of the shared boundary, terrain connectivity, road and water system barriers, and the similarity of historical utilization status. The edge weight value ranges from [0,1]. Among them, the closer the map features are, the longer the shared boundary is, the higher the terrain connectivity, and the more similar the historical utilization status, the greater the edge weight; the higher the degree of road or water system barrier, the smaller the edge weight.

[0223] Specifically, based on historically verified abandoned land parcel records and currently suspected continuously abandoned land parcels, the abandonment diffusion front is identified in the cultivated land parcel spatial neighborhood relationship map. The abandonment diffusion front refers to the boundary region formed by cultivated land parcels in the cultivated land parcel spatial neighborhood relationship map that are directly adjacent to historically verified abandoned land parcel records or currently suspected continuously abandoned land parcels through edges, and whose edge weights are greater than a preset diffusion adjacency threshold. The preset diffusion adjacency threshold is determined based on the historical distribution of edge weights between historically verified abandoned land parcel records and their adjacent abandoned land parcels, specifically taking the median of this historical distribution. For example, the preset diffusion adjacency threshold is set to 0.6. Cultivated land parcels located within the abandonment diffusion front and showing a decrease in phenological activity are marked as spatial diffusion risk sensitive parcels. The aforementioned spatial diffusion risk sensitive patches refer to cultivated land patches that have not yet formed suspected phenological breakpoints, but are located close to historically verified abandoned patches or currently continuously abandoned suspected patches, and whose own phenological activity has declined, thus posing a risk of subsequent abandonment.

[0224] In step S43, a spatial neighborhood relationship graph of cultivated land parcels is constructed, using cultivated land parcels as nodes and spatial adjacency relationships or distances less than a preset neighborhood radius between parcels as edges. The edge weight calculation comprehensively considers factors such as distance between parcels, shared boundary length, terrain connectivity, road and water system barriers, and similarity of historical utilization status. This spatial neighborhood relationship graph can identify the frontier of abandoned land spread and generate a list of abandoned land risk warnings for cultivated land parcels located at the frontier and showing a decline in phenological activity.

[0225] Specifically, the phenological activity is a dimensionless index characterizing the activity level of current crop phenological processes in cultivated plots. It is obtained by weighting the ratio of peak vegetation cover in the core area to the baseline of peak vegetation cover in the core area, the ratio of tillage disturbance frequency to the baseline of tillage disturbance frequency, and the degree of agreement between phenological stages and phenological state probability windows. It is uniformly mapped to the [0,1] interval, with higher values ​​indicating more active crop phenological processes. Continuous decay of phenological activity refers to a sustained decrease in phenological activity across consecutive observation periods.

[0226] Specifically, risk assessments are conducted on cultivated plots that have not yet formed a true fallow breakpoint but have shown continuous decline in phenological activity, a continuous expansion of the deviation of the plot-level phenological benchmark fingerprint, a decrease in the peak vegetation cover in the core area, or a reduction in the frequency of tillage disturbance.

[0227] Specifically, when a cultivated plot simultaneously meets both the conditions of continuous decline in its own phenological activity and being located at the forefront of fallow spread, the corresponding cultivated plot will be classified as a high-risk warning level; when a cultivated plot only meets one of the conditions of continuous decline in its own phenological activity or being located at the forefront of fallow spread, the corresponding cultivated plot will be classified as a medium-risk warning level.

[0228] Specifically, when the decline in phenological activity is highly correlated with regional drought, floods, extreme low temperatures, or other climate anomalies, climate disturbance markers are recorded, and the corresponding cultivated land patches are temporarily deferred from being directly identified as abandoned risk patches.

[0229] S44: Output the final abandonment status identification result based on the review results, the breakpoint reversibility judgment results, and the abandonment risk warning list.

[0230] The final result of the abandoned state identification needs to include the state type, spatial range, credibility, false breakpoint cause, reversibility state, and priority of manual verification, so as to carry out subsequent manual verification and state tracking.

[0231] Specifically, based on the breakpoint credible potential energy, pseudo-breakpoint verification results, breakpoint reversibility judgment results, and abandonment risk warning list, the final abandonment status identification result is generated.

[0232] Specifically, the final fallow status identification results include cultivated land patch number, fallow status type, suspected fallow occurrence period, fallow duration, fallow spatial range, fallow confidence, breakpoint confidence potential energy, deviation amplitude of patch-level phenological benchmark fingerprint, reversibility judgment result, pseudo-breakpoint inducing factor marker, spatial diffusion risk marker, climate interference marker, priority of manual verification, and breakpoint interpretation chain. Among these, the suspected fallow occurrence period is represented by the observation period or observation date; the fallow duration is represented by the number of days or consecutive observation periods; the fallow spatial range is represented by the patch area or spatial vector range; and the fallow confidence, breakpoint confidence potential energy, and deviation amplitude of patch-level phenological benchmark fingerprint are all dimensionless discriminant indicators. The fallow confidence corresponds positively to the actual fallow probability. The fallow confidence adopts the same... The normalized value corresponding to a monotonically inverse inverse; The lower the value, the more credible the abandonment is.

[0233] Specifically, the types of fallow status include normal cultivation, seasonal fallow, weakly recoverable cultivation, continuous fallow, potential energy pending verification pseudo-breakpoints, and fallow risk warning.

[0234] Specifically, when the abandonment status type is continuous abandonment, the priority of manual verification is set to high priority, and the manual verification task is output to the farmland supervision platform; when the abandonment status type is seasonal fallow or weakly recoverable cultivation, the corresponding farmland patch is written into the automatic tracking queue of the next image; when the abandonment status type is a pseudo-breakpoint with potential energy pending verification, the pseudo-breakpoint cause marker is output to the image processing module and the patch boundary maintenance module; when the abandonment status type is an abandonment risk warning patch, the corresponding farmland patch is written into the risk warning list.

[0235] Specifically, the farmland monitoring platform generates a map patch verification list, a risk warning list, and a status statistical report based on the final abandonment status identification results. The map patch verification list is used for on-site verification, the risk warning list is used for key monitoring in the next observation cycle, and the status statistical report is used for farmland outflow status analysis.

[0236] To further illustrate the above identification process, let's take a cultivated land parcel to be monitored as an example. The historical normal cultivation periods of this cultivated land parcel can form a stable phenological rhythm timeline, with phenological state changes corresponding to the greening, vigorous growth, harvesting, and tillage processes within the corresponding observation periods. After processing the current multi-period images, if the phenological state of the core area changes from a normal cultivation state to a weak cultivation state or a state without cultivation disturbance in consecutive observation periods, and at the same time, the vegetation cover, bare soil exposure cycle, or tillage disturbance in the core area continuously deviates from the parcel-level phenological baseline fingerprint, then a candidate phenological breakpoint is formed.

[0237] When both the spatial feature view and the temporal phenological view give a consistent judgment on the state interruption in the core area, the reliability potential energy of the interruption is calculated by combining the spatial relationships between the candidate phenological breakpoints and the plot boundaries, image tile seams, adjacent object interference areas, and cross-period boundary drift areas, and suspected abandoned phenological breakpoints are screened. For the screened suspected abandoned phenological breakpoints, reverse verification is performed by image tile seams, boundary drift, adjacent object interference, and contour prediction; if the anomaly continues to be located in the core area and cannot be explained by the above-mentioned pseudo-breakpoint causes, the abandoned judgment is retained. If a replanting signal appears in subsequent observation periods, the corresponding cultivated plot is determined as seasonal fallow or recoverable weak cultivation based on the reversibility judgment result; if the phenological interruption continues and no replanting signal appears, the corresponding cultivated plot is determined as a suspected plot of continuous abandoned land. Thus, continuous abandoned land, recoverable state, and pseudo-anomalies caused by image and boundary factors can be distinguished in the same processing flow.

[0238] Example 2:

[0239] This embodiment, based on Embodiment 1, provides a temporal phenological breakpoint identification system for farmland abandonment, used to implement the aforementioned method for identifying temporal phenological breakpoints in farmland abandonment. The system can identify suspected abandonment phenological breakpoints using farmland patches as the basic object, even under conditions of registration errors, patch boundary drift, image slice seam interference, influence from adjacent objects, and regional differences in multi-period remote sensing images. It then outputs the abandonment status type, breakpoint interpretation chain, and risk warning results.

[0240] like Figure 4 The diagram shown is a functional module diagram of a time-series phenological breakpoint identification system for abandoned farmland. The system includes a map feature and time-series unit generation module, a spatial partitioning and phenological state modeling module, a breakpoint credibility potential energy and reversibility judgment module, and a verification output and risk warning module.

[0241] In this system embodiment, the client can be used to display the final abandoned status identification results, breakpoint interpretation chains, map patch verification lists, and risk warning lists. Thus, this invention can systematically complete the entire processing flow from image input, model recognition, breakpoint determination to regulatory result output.

[0242] The image patch benchmark and time series unit generation module is used to construct image patch-level phenological benchmark fingerprints based on historical normal farming periods, and then aggregate the current image slices according to the cultivated map patches to form time series image units.

[0243] The spatial zoning and phenological state modeling module is used to construct a spatial zoning confidence map based on the time-series image unit, and then use the spatial zoning confidence map to constrain cross-view phenological state modeling to form a patch-level time-series phenological state sequence.

[0244] The breakpoint credibility potential energy and reversibility judgment module is used to extract candidate phenological breakpoints based on the patch-level time-series phenological state sequence, then calculate the breakpoint credibility potential energy of the candidate phenological breakpoints, then screen suspected abandoned phenological breakpoints based on the breakpoint credibility potential energy, and finally judge the breakpoint reversibility to form the breakpoint type.

[0245] The review output and risk warning module is used to perform reverse review of the suspected abandoned phenological breakpoints based on the breakpoint type, and then form the final abandoned state identification result based on the review result, and subsequently generate a risk warning list.

[0246] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, combinations, or omissions made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying time-series phenological breakpoints in the state of abandoned farmland, characterized in that, include: Based on historical normal farming periods, a patch-level phenological benchmark fingerprint is constructed, and then the current image slices are aggregated according to the cultivated patch to form a time-series image unit. Spatial partition confidence maps are constructed based on temporal image units, and then the spatial partition confidence maps are used to constrain cross-view phenological state modeling to form a patch-level temporal phenological state sequence. Candidate phenological breakpoints are extracted based on the time-series phenological state sequence at the map patch level. Then, the breakpoint reliability potential energy of the candidate phenological breakpoints is calculated. Subsequently, suspected abandoned phenological breakpoints are screened based on the breakpoint reliability potential energy. Finally, the breakpoint reversibility is determined to form the breakpoint type. Based on the breakpoint type, a reverse verification of the pseudo-breakpoint causes is performed on the suspected abandoned phenological breakpoints. Then, based on the verification results, the final abandoned status identification result is formed, and a risk warning list is generated.

2. The method for identifying time-series phenological breakpoints in the state of abandoned farmland according to claim 1, characterized in that, The method for constructing a patch-level phenological benchmark fingerprint based on historical normal farming periods includes: For the same cultivated land patch, the occurrence time windows of each phenological stage are extracted from the historical normal cultivation period to generate a historical phenological rhythm timeline; Based on the distribution of observation dates corresponding to each phenological stage according to the historical normal farming period, the probability window of phenological state is determined. The patch-level phenological baseline fingerprint is formed by combining the historical phenological rhythm timeline with the phenological state probability window.

3. The method for identifying time-series phenological breakpoints in the state of abandoned farmland according to claim 2, characterized in that, The probability window for determining phenological states includes: For any of the phenological stages, the distribution of observation dates corresponding to the phenological stages in the historical normal farming periods is statistically analyzed; Determine the date corresponding to the 10th percentile based on the observed date distribution; The date corresponding to the 90th percentile is determined based on the same observed date distribution; The date range between the date corresponding to the 10th percentile and the date corresponding to the 90th percentile is determined as the phenological state probability window.

4. The method for identifying time-series phenological breakpoints in the state of abandoned farmland according to claim 1, characterized in that, The process of aggregating current image slices according to cultivated land patches to form time-series image units includes: Establish a parcel semantic anchoring relationship for each of the current image slices; The farmland patch number corresponding to the current image slice is determined based on the semantic anchoring relationship of the land parcel; The current image slices of the same cultivated land patch are aggregated using the cultivated land patch number as the primary key; The current image slices, arranged and aggregated according to the observation period, form the time-series image unit.

5. The method for identifying time-series phenological breakpoints in the state of abandoned farmland according to claim 4, characterized in that, The establishment of semantic anchoring relationships for land parcels includes: The current image slice is spatially overlaid with the boundary data of cultivated land patches to determine the number of cultivated land patches covered by the current image slice. The relative spatial position of the current image slice in the corresponding cultivated land patch is determined based on the spatial overlay result; The interference relationship between adjacent objects is determined based on the surrounding features of the cultivated land plot boundary. The cultivated land parcel number is associated with the current image slice, the relative spatial position is recorded, and then the interference relationship of adjacent objects is written to form the semantic anchoring relationship of the land parcel.

6. The method for identifying time-series phenological breakpoints in the state of abandoned farmland according to claim 1, characterized in that, The construction of spatial partition confidence maps based on temporal image units includes: Extract the basic image feature set from the temporal image unit; Based on the aforementioned basic image feature set, a fine segmentation result of the land parcel is generated; The spatial region to which each spatial unit belongs is determined based on the detailed internal subdivision results of the plot. The confidence level of spatial partitioning is determined based on the position of the spatial region relative to the boundary of cultivated land patches. The spatial partition confidence scores are arranged according to geographical location to generate the spatial partition confidence map.

7. The method for identifying time-series phenological breakpoints in the state of abandoned farmland according to claim 6, characterized in that, The cross-view phenological state modeling includes: A spatial feature view constrained by the spatial partition confidence map is constructed based on the fine segmentation results within the plot. The core region is determined based on the spatial partition confidence map, and the core region image sequence corresponding to the core region is extracted from the temporal image unit; Based on the patch-level phenological benchmark fingerprint, the image sequence of the core area is compared to construct a time-series phenological view; When the spatial feature view indicates that the core area has a state interruption during the target observation period and the time-series phenological view determines that the state interruption deviates from the patch-level phenological benchmark fingerprint, a suspected phenological anomaly marker is generated. The suspected phenological anomaly markers are recorded according to the observation period to form the patch-level time-series phenological state sequence.

8. The method for identifying time-series phenological breakpoints in the state of abandoned farmland according to claim 1, characterized in that, The calculation of the reliable potential energy at the breakpoint includes: The degree of overlap between the candidate phenological breakpoint and the land parcel boundary buffer zone is obtained, and the degree of overlap between the candidate phenological breakpoint and the image slice seam buffer zone is obtained, thereby determining the intensity of the false breakpoint induction. The support strength of the true breakpoint is determined based on the sustained significance of the candidate phenological breakpoints within the core region, and then the support strength of the true breakpoints is adjusted based on the sustained deviation of the candidate phenological breakpoints from the patch-level phenological benchmark fingerprint. The false breakpoint induction strength is used as the potential increase, and the adjusted real breakpoint support strength is used as the potential decrease to generate the breakpoint credible potential energy. Candidate phenological breakpoints whose breakpoint credible potential energy is lower than a preset breakpoint potential energy threshold are identified as suspected abandoned phenological breakpoints.

9. The method for identifying time-series phenological breakpoints in the state of abandoned farmland according to claim 8, characterized in that, The determination of breakpoint reversibility forms the breakpoint type, including: Extract a set of reversibility features from the observation periods before and after the suspected phenological breakpoint of abandoned land; Based on the set of reversible features, determine whether the corresponding cultivated land patch retains the plot maintenance features; When the maintenance characteristics of the land parcel are preserved, the suspected abandoned phenological breakpoints are determined to be reversible breakpoints; When the maintenance characteristics of the land parcel are not preserved and irreversible degradation characteristics appear, the suspected abandoned phenological breakpoint is determined to be an irreversible breakpoint. The breakpoint type is formed by combining the reversible breakpoint or the irreversible breakpoint obtained from the determination with the abnormal duration period; The breakpoint type is used as input for the reverse verification of pseudo-breakpoint causes.

10. The method for identifying time-series phenological breakpoints in the state of abandoned farmland according to claim 9, characterized in that, The reverse verification of the pseudo-breakpoint causes includes: First, perform image slice seam verification on the suspected abandoned phenological breakpoints; Perform boundary drift verification on the retained suspected abandoned phenological breakpoints; For the suspected abandoned phenological breakpoints that are retained again, perform a verification of interference from adjacent objects; Perform a reverse verification of the profile prediction on the suspected abandoned phenological breakpoints that are retained again to obtain the verification results; The final abandonment status identification result is formed based on the verification results; A spatial neighborhood relationship graph of cultivated land patches is constructed using cultivated land patches as nodes. Based on the final abandonment status identification result, the abandonment diffusion front is identified in the spatial neighborhood relationship graph of cultivated land patches. The risk warning list is generated based on the abandonment diffusion front.

11. A system for identifying time-series phenological breakpoints in the state of abandoned farmland, characterized in that, include: The image patch benchmark and time series unit generation module is used to construct image patch-level phenological benchmark fingerprints based on historical normal farming periods, and then aggregate the current image slices according to the cultivated map patches to form time series image units. The spatial zoning and phenological state modeling module is used to construct a spatial zoning confidence map based on the time-series image unit, and then use the spatial zoning confidence map to constrain cross-view phenological state modeling to form a patch-level time-series phenological state sequence. The breakpoint credibility potential energy and reversibility judgment module is used to extract candidate phenological breakpoints based on the patch-level time-series phenological state sequence, then calculate the breakpoint credibility potential energy of the candidate phenological breakpoints, then screen suspected abandoned phenological breakpoints based on the breakpoint credibility potential energy, and finally judge the breakpoint reversibility to form the breakpoint type. The review output and risk warning module is used to perform reverse review of the suspected abandoned phenological breakpoints based on the breakpoint type, and then form the final abandoned state identification result based on the review result, and subsequently generate a risk warning list.

Images